Calibration of word posterior estimation in confusion networks for keyword search

Word posterior probability has been widely used as the confidence estimation of automatic speech recognition (ASR) systems and has been proved to be quite effective in related applications such as keyword search. However, word posterior probability tends to overestimate the true probability of a hypothesis, as it is computed on a subset of the total hypothesis space. In this paper, we show that a more accurate estimation of posterior can be obtained by using a calibration method based on the conditional random field (CRF) model. By using calibrated posterior estimation for keyword search task, we obtain a maximum absolute gain of 1.15% for single-word keyword search on the maximum term-weighted value (MTWV) metric on the OpenKWS14 Tamil dataset.